AI Model Temperaments: Why Coding Defaults Matter More Than Benchmarks
Updated
AI coding models are not interchangeable capability engines. Given the same vague prompt, one model will invent three new abstractions before you finish reading its plan, while another stops and asks which of two paths you meant. Same capability tier, different reflexes. Those reflexes compound across every session and shape the long-term coherence of your codebase.
What "Temperament" Means for a Model
Model temperament is not personality in a mystical sense. It refers to working defaults: how a model behaves when the prompt leaves room for interpretation. Key axes include whether it asks or assumes, whether diffs stay narrow or spread, and whether it leans on existing conventions or generates new structure.
Why Defaults Dominate
A model resets every session. Whatever discipline you coached into it last time is gone by the next context window. Unlike a human colleague who absorbs your conventions over months, a model returns to its factory defaults each time. This makes the default temperament the thing to select for, because it is what you get every single time.
Q&A
Do AI coding models really have different temperaments?
Yes. Given an ambiguous task, different models exhibit distinct default behaviors: some ask clarifying questions, others pick a path and run. Some keep changes narrow, others refactor broadly. These are not random variations but consistent patterns in how each model handles uncertainty and tradeoffs.
Why can't you just fix a model's temperament with a system prompt?
System prompts and standing instructions help at the margins, but they only cover the cases you anticipated. Most real coding decisions fall into gaps between your rules, and in those gaps the model falls back on its trained defaults. The base temperament is the backstop for everything you did not write a rule for.
What does the sailboat keel metaphor mean in this context?
A racing sailboat's keel is dead weight that adds no speed, but it prevents capsize when the wind gets serious. A disciplined model works the same way: it may feel like it slows you down by asking questions and preserving structure, but it converts your speed into maintainable software instead of letting the codebase tip over under real load.
Is a disciplined model always the right choice?
No. If your own temperament already over-supplies caution, polish, and structure, a disciplined model just adds more of the same. In that case, you need a fast, inventive partner that supplies momentum and a tolerance for rough edges. The principle is to choose for the imbalance, whichever direction it runs.